Why is the O3 Reasoning Model Underperforming on API?
Asked by
CuriousCoder99
I've been experimenting with some complex prompts to encourage the O3-mini model to engage in reasoning when using the API, but it just spits out answers without any proper reasoning. On the other hand, the O3 model in ChatGPT takes its time and really dives deep into reasoning, even utilizing Python functions and handling images effectively. What are the main factors that would help bring that level of reasoning to the API? It's also frustrating that O3 seems to be only available for internal use while we can only access O3-mini through the API. Has anyone else experienced this issue?
Answers
It’s actually not suggested to try to "force reasoning" with the O3 model since it normally does that natively. Trying to prompt it for reasoning might even lower its performance. Also, O3 is available in the API, so it seems like there might be some confusion there. Have you checked out different prompt examples for it?
Reply to this answer
I think you might need to be on a higher tier of the API to access those features. But you’re correct, OpenAI has released O3 to the API. It's meant to reason on its own, but I totally get your frustration if it isn’t performing well. Did your tests show any improvement when adjusting the prompts? Also, implementing sandbox function calls could help improve image handling.
Reply to this answer
Add your answer
Answers are reviewed before publication.
CuriousCoder99: Thanks for the input! I’m curious about how to implement those sandbox functions better. Do you have any specific examples or suggestions?